Loading...
1#include <trace/syscall.h>
2#include <trace/events/syscalls.h>
3#include <linux/slab.h>
4#include <linux/kernel.h>
5#include <linux/module.h> /* for MODULE_NAME_LEN via KSYM_SYMBOL_LEN */
6#include <linux/ftrace.h>
7#include <linux/perf_event.h>
8#include <asm/syscall.h>
9
10#include "trace_output.h"
11#include "trace.h"
12
13static DEFINE_MUTEX(syscall_trace_lock);
14static int sys_refcount_enter;
15static int sys_refcount_exit;
16static DECLARE_BITMAP(enabled_enter_syscalls, NR_syscalls);
17static DECLARE_BITMAP(enabled_exit_syscalls, NR_syscalls);
18
19static int syscall_enter_register(struct ftrace_event_call *event,
20 enum trace_reg type, void *data);
21static int syscall_exit_register(struct ftrace_event_call *event,
22 enum trace_reg type, void *data);
23
24static int syscall_enter_define_fields(struct ftrace_event_call *call);
25static int syscall_exit_define_fields(struct ftrace_event_call *call);
26
27static struct list_head *
28syscall_get_enter_fields(struct ftrace_event_call *call)
29{
30 struct syscall_metadata *entry = call->data;
31
32 return &entry->enter_fields;
33}
34
35struct trace_event_functions enter_syscall_print_funcs = {
36 .trace = print_syscall_enter,
37};
38
39struct trace_event_functions exit_syscall_print_funcs = {
40 .trace = print_syscall_exit,
41};
42
43struct ftrace_event_class event_class_syscall_enter = {
44 .system = "syscalls",
45 .reg = syscall_enter_register,
46 .define_fields = syscall_enter_define_fields,
47 .get_fields = syscall_get_enter_fields,
48 .raw_init = init_syscall_trace,
49};
50
51struct ftrace_event_class event_class_syscall_exit = {
52 .system = "syscalls",
53 .reg = syscall_exit_register,
54 .define_fields = syscall_exit_define_fields,
55 .fields = LIST_HEAD_INIT(event_class_syscall_exit.fields),
56 .raw_init = init_syscall_trace,
57};
58
59extern struct syscall_metadata *__start_syscalls_metadata[];
60extern struct syscall_metadata *__stop_syscalls_metadata[];
61
62static struct syscall_metadata **syscalls_metadata;
63
64#ifndef ARCH_HAS_SYSCALL_MATCH_SYM_NAME
65static inline bool arch_syscall_match_sym_name(const char *sym, const char *name)
66{
67 /*
68 * Only compare after the "sys" prefix. Archs that use
69 * syscall wrappers may have syscalls symbols aliases prefixed
70 * with "SyS" instead of "sys", leading to an unwanted
71 * mismatch.
72 */
73 return !strcmp(sym + 3, name + 3);
74}
75#endif
76
77static __init struct syscall_metadata *
78find_syscall_meta(unsigned long syscall)
79{
80 struct syscall_metadata **start;
81 struct syscall_metadata **stop;
82 char str[KSYM_SYMBOL_LEN];
83
84
85 start = __start_syscalls_metadata;
86 stop = __stop_syscalls_metadata;
87 kallsyms_lookup(syscall, NULL, NULL, NULL, str);
88
89 if (arch_syscall_match_sym_name(str, "sys_ni_syscall"))
90 return NULL;
91
92 for ( ; start < stop; start++) {
93 if ((*start)->name && arch_syscall_match_sym_name(str, (*start)->name))
94 return *start;
95 }
96 return NULL;
97}
98
99static struct syscall_metadata *syscall_nr_to_meta(int nr)
100{
101 if (!syscalls_metadata || nr >= NR_syscalls || nr < 0)
102 return NULL;
103
104 return syscalls_metadata[nr];
105}
106
107enum print_line_t
108print_syscall_enter(struct trace_iterator *iter, int flags,
109 struct trace_event *event)
110{
111 struct trace_seq *s = &iter->seq;
112 struct trace_entry *ent = iter->ent;
113 struct syscall_trace_enter *trace;
114 struct syscall_metadata *entry;
115 int i, ret, syscall;
116
117 trace = (typeof(trace))ent;
118 syscall = trace->nr;
119 entry = syscall_nr_to_meta(syscall);
120
121 if (!entry)
122 goto end;
123
124 if (entry->enter_event->event.type != ent->type) {
125 WARN_ON_ONCE(1);
126 goto end;
127 }
128
129 ret = trace_seq_printf(s, "%s(", entry->name);
130 if (!ret)
131 return TRACE_TYPE_PARTIAL_LINE;
132
133 for (i = 0; i < entry->nb_args; i++) {
134 /* parameter types */
135 if (trace_flags & TRACE_ITER_VERBOSE) {
136 ret = trace_seq_printf(s, "%s ", entry->types[i]);
137 if (!ret)
138 return TRACE_TYPE_PARTIAL_LINE;
139 }
140 /* parameter values */
141 ret = trace_seq_printf(s, "%s: %lx%s", entry->args[i],
142 trace->args[i],
143 i == entry->nb_args - 1 ? "" : ", ");
144 if (!ret)
145 return TRACE_TYPE_PARTIAL_LINE;
146 }
147
148 ret = trace_seq_putc(s, ')');
149 if (!ret)
150 return TRACE_TYPE_PARTIAL_LINE;
151
152end:
153 ret = trace_seq_putc(s, '\n');
154 if (!ret)
155 return TRACE_TYPE_PARTIAL_LINE;
156
157 return TRACE_TYPE_HANDLED;
158}
159
160enum print_line_t
161print_syscall_exit(struct trace_iterator *iter, int flags,
162 struct trace_event *event)
163{
164 struct trace_seq *s = &iter->seq;
165 struct trace_entry *ent = iter->ent;
166 struct syscall_trace_exit *trace;
167 int syscall;
168 struct syscall_metadata *entry;
169 int ret;
170
171 trace = (typeof(trace))ent;
172 syscall = trace->nr;
173 entry = syscall_nr_to_meta(syscall);
174
175 if (!entry) {
176 trace_seq_printf(s, "\n");
177 return TRACE_TYPE_HANDLED;
178 }
179
180 if (entry->exit_event->event.type != ent->type) {
181 WARN_ON_ONCE(1);
182 return TRACE_TYPE_UNHANDLED;
183 }
184
185 ret = trace_seq_printf(s, "%s -> 0x%lx\n", entry->name,
186 trace->ret);
187 if (!ret)
188 return TRACE_TYPE_PARTIAL_LINE;
189
190 return TRACE_TYPE_HANDLED;
191}
192
193extern char *__bad_type_size(void);
194
195#define SYSCALL_FIELD(type, name) \
196 sizeof(type) != sizeof(trace.name) ? \
197 __bad_type_size() : \
198 #type, #name, offsetof(typeof(trace), name), \
199 sizeof(trace.name), is_signed_type(type)
200
201static
202int __set_enter_print_fmt(struct syscall_metadata *entry, char *buf, int len)
203{
204 int i;
205 int pos = 0;
206
207 /* When len=0, we just calculate the needed length */
208#define LEN_OR_ZERO (len ? len - pos : 0)
209
210 pos += snprintf(buf + pos, LEN_OR_ZERO, "\"");
211 for (i = 0; i < entry->nb_args; i++) {
212 pos += snprintf(buf + pos, LEN_OR_ZERO, "%s: 0x%%0%zulx%s",
213 entry->args[i], sizeof(unsigned long),
214 i == entry->nb_args - 1 ? "" : ", ");
215 }
216 pos += snprintf(buf + pos, LEN_OR_ZERO, "\"");
217
218 for (i = 0; i < entry->nb_args; i++) {
219 pos += snprintf(buf + pos, LEN_OR_ZERO,
220 ", ((unsigned long)(REC->%s))", entry->args[i]);
221 }
222
223#undef LEN_OR_ZERO
224
225 /* return the length of print_fmt */
226 return pos;
227}
228
229static int set_syscall_print_fmt(struct ftrace_event_call *call)
230{
231 char *print_fmt;
232 int len;
233 struct syscall_metadata *entry = call->data;
234
235 if (entry->enter_event != call) {
236 call->print_fmt = "\"0x%lx\", REC->ret";
237 return 0;
238 }
239
240 /* First: called with 0 length to calculate the needed length */
241 len = __set_enter_print_fmt(entry, NULL, 0);
242
243 print_fmt = kmalloc(len + 1, GFP_KERNEL);
244 if (!print_fmt)
245 return -ENOMEM;
246
247 /* Second: actually write the @print_fmt */
248 __set_enter_print_fmt(entry, print_fmt, len + 1);
249 call->print_fmt = print_fmt;
250
251 return 0;
252}
253
254static void free_syscall_print_fmt(struct ftrace_event_call *call)
255{
256 struct syscall_metadata *entry = call->data;
257
258 if (entry->enter_event == call)
259 kfree(call->print_fmt);
260}
261
262static int syscall_enter_define_fields(struct ftrace_event_call *call)
263{
264 struct syscall_trace_enter trace;
265 struct syscall_metadata *meta = call->data;
266 int ret;
267 int i;
268 int offset = offsetof(typeof(trace), args);
269
270 ret = trace_define_field(call, SYSCALL_FIELD(int, nr), FILTER_OTHER);
271 if (ret)
272 return ret;
273
274 for (i = 0; i < meta->nb_args; i++) {
275 ret = trace_define_field(call, meta->types[i],
276 meta->args[i], offset,
277 sizeof(unsigned long), 0,
278 FILTER_OTHER);
279 offset += sizeof(unsigned long);
280 }
281
282 return ret;
283}
284
285static int syscall_exit_define_fields(struct ftrace_event_call *call)
286{
287 struct syscall_trace_exit trace;
288 int ret;
289
290 ret = trace_define_field(call, SYSCALL_FIELD(int, nr), FILTER_OTHER);
291 if (ret)
292 return ret;
293
294 ret = trace_define_field(call, SYSCALL_FIELD(long, ret),
295 FILTER_OTHER);
296
297 return ret;
298}
299
300void ftrace_syscall_enter(void *ignore, struct pt_regs *regs, long id)
301{
302 struct syscall_trace_enter *entry;
303 struct syscall_metadata *sys_data;
304 struct ring_buffer_event *event;
305 struct ring_buffer *buffer;
306 int size;
307 int syscall_nr;
308
309 syscall_nr = syscall_get_nr(current, regs);
310 if (syscall_nr < 0)
311 return;
312 if (!test_bit(syscall_nr, enabled_enter_syscalls))
313 return;
314
315 sys_data = syscall_nr_to_meta(syscall_nr);
316 if (!sys_data)
317 return;
318
319 size = sizeof(*entry) + sizeof(unsigned long) * sys_data->nb_args;
320
321 event = trace_current_buffer_lock_reserve(&buffer,
322 sys_data->enter_event->event.type, size, 0, 0);
323 if (!event)
324 return;
325
326 entry = ring_buffer_event_data(event);
327 entry->nr = syscall_nr;
328 syscall_get_arguments(current, regs, 0, sys_data->nb_args, entry->args);
329
330 if (!filter_current_check_discard(buffer, sys_data->enter_event,
331 entry, event))
332 trace_current_buffer_unlock_commit(buffer, event, 0, 0);
333}
334
335void ftrace_syscall_exit(void *ignore, struct pt_regs *regs, long ret)
336{
337 struct syscall_trace_exit *entry;
338 struct syscall_metadata *sys_data;
339 struct ring_buffer_event *event;
340 struct ring_buffer *buffer;
341 int syscall_nr;
342
343 syscall_nr = syscall_get_nr(current, regs);
344 if (syscall_nr < 0)
345 return;
346 if (!test_bit(syscall_nr, enabled_exit_syscalls))
347 return;
348
349 sys_data = syscall_nr_to_meta(syscall_nr);
350 if (!sys_data)
351 return;
352
353 event = trace_current_buffer_lock_reserve(&buffer,
354 sys_data->exit_event->event.type, sizeof(*entry), 0, 0);
355 if (!event)
356 return;
357
358 entry = ring_buffer_event_data(event);
359 entry->nr = syscall_nr;
360 entry->ret = syscall_get_return_value(current, regs);
361
362 if (!filter_current_check_discard(buffer, sys_data->exit_event,
363 entry, event))
364 trace_current_buffer_unlock_commit(buffer, event, 0, 0);
365}
366
367int reg_event_syscall_enter(struct ftrace_event_call *call)
368{
369 int ret = 0;
370 int num;
371
372 num = ((struct syscall_metadata *)call->data)->syscall_nr;
373 if (WARN_ON_ONCE(num < 0 || num >= NR_syscalls))
374 return -ENOSYS;
375 mutex_lock(&syscall_trace_lock);
376 if (!sys_refcount_enter)
377 ret = register_trace_sys_enter(ftrace_syscall_enter, NULL);
378 if (!ret) {
379 set_bit(num, enabled_enter_syscalls);
380 sys_refcount_enter++;
381 }
382 mutex_unlock(&syscall_trace_lock);
383 return ret;
384}
385
386void unreg_event_syscall_enter(struct ftrace_event_call *call)
387{
388 int num;
389
390 num = ((struct syscall_metadata *)call->data)->syscall_nr;
391 if (WARN_ON_ONCE(num < 0 || num >= NR_syscalls))
392 return;
393 mutex_lock(&syscall_trace_lock);
394 sys_refcount_enter--;
395 clear_bit(num, enabled_enter_syscalls);
396 if (!sys_refcount_enter)
397 unregister_trace_sys_enter(ftrace_syscall_enter, NULL);
398 mutex_unlock(&syscall_trace_lock);
399}
400
401int reg_event_syscall_exit(struct ftrace_event_call *call)
402{
403 int ret = 0;
404 int num;
405
406 num = ((struct syscall_metadata *)call->data)->syscall_nr;
407 if (WARN_ON_ONCE(num < 0 || num >= NR_syscalls))
408 return -ENOSYS;
409 mutex_lock(&syscall_trace_lock);
410 if (!sys_refcount_exit)
411 ret = register_trace_sys_exit(ftrace_syscall_exit, NULL);
412 if (!ret) {
413 set_bit(num, enabled_exit_syscalls);
414 sys_refcount_exit++;
415 }
416 mutex_unlock(&syscall_trace_lock);
417 return ret;
418}
419
420void unreg_event_syscall_exit(struct ftrace_event_call *call)
421{
422 int num;
423
424 num = ((struct syscall_metadata *)call->data)->syscall_nr;
425 if (WARN_ON_ONCE(num < 0 || num >= NR_syscalls))
426 return;
427 mutex_lock(&syscall_trace_lock);
428 sys_refcount_exit--;
429 clear_bit(num, enabled_exit_syscalls);
430 if (!sys_refcount_exit)
431 unregister_trace_sys_exit(ftrace_syscall_exit, NULL);
432 mutex_unlock(&syscall_trace_lock);
433}
434
435int init_syscall_trace(struct ftrace_event_call *call)
436{
437 int id;
438 int num;
439
440 num = ((struct syscall_metadata *)call->data)->syscall_nr;
441 if (num < 0 || num >= NR_syscalls) {
442 pr_debug("syscall %s metadata not mapped, disabling ftrace event\n",
443 ((struct syscall_metadata *)call->data)->name);
444 return -ENOSYS;
445 }
446
447 if (set_syscall_print_fmt(call) < 0)
448 return -ENOMEM;
449
450 id = trace_event_raw_init(call);
451
452 if (id < 0) {
453 free_syscall_print_fmt(call);
454 return id;
455 }
456
457 return id;
458}
459
460unsigned long __init __weak arch_syscall_addr(int nr)
461{
462 return (unsigned long)sys_call_table[nr];
463}
464
465int __init init_ftrace_syscalls(void)
466{
467 struct syscall_metadata *meta;
468 unsigned long addr;
469 int i;
470
471 syscalls_metadata = kcalloc(NR_syscalls, sizeof(*syscalls_metadata),
472 GFP_KERNEL);
473 if (!syscalls_metadata) {
474 WARN_ON(1);
475 return -ENOMEM;
476 }
477
478 for (i = 0; i < NR_syscalls; i++) {
479 addr = arch_syscall_addr(i);
480 meta = find_syscall_meta(addr);
481 if (!meta)
482 continue;
483
484 meta->syscall_nr = i;
485 syscalls_metadata[i] = meta;
486 }
487
488 return 0;
489}
490core_initcall(init_ftrace_syscalls);
491
492#ifdef CONFIG_PERF_EVENTS
493
494static DECLARE_BITMAP(enabled_perf_enter_syscalls, NR_syscalls);
495static DECLARE_BITMAP(enabled_perf_exit_syscalls, NR_syscalls);
496static int sys_perf_refcount_enter;
497static int sys_perf_refcount_exit;
498
499static void perf_syscall_enter(void *ignore, struct pt_regs *regs, long id)
500{
501 struct syscall_metadata *sys_data;
502 struct syscall_trace_enter *rec;
503 struct hlist_head *head;
504 int syscall_nr;
505 int rctx;
506 int size;
507
508 syscall_nr = syscall_get_nr(current, regs);
509 if (!test_bit(syscall_nr, enabled_perf_enter_syscalls))
510 return;
511
512 sys_data = syscall_nr_to_meta(syscall_nr);
513 if (!sys_data)
514 return;
515
516 /* get the size after alignment with the u32 buffer size field */
517 size = sizeof(unsigned long) * sys_data->nb_args + sizeof(*rec);
518 size = ALIGN(size + sizeof(u32), sizeof(u64));
519 size -= sizeof(u32);
520
521 if (WARN_ONCE(size > PERF_MAX_TRACE_SIZE,
522 "perf buffer not large enough"))
523 return;
524
525 rec = (struct syscall_trace_enter *)perf_trace_buf_prepare(size,
526 sys_data->enter_event->event.type, regs, &rctx);
527 if (!rec)
528 return;
529
530 rec->nr = syscall_nr;
531 syscall_get_arguments(current, regs, 0, sys_data->nb_args,
532 (unsigned long *)&rec->args);
533
534 head = this_cpu_ptr(sys_data->enter_event->perf_events);
535 perf_trace_buf_submit(rec, size, rctx, 0, 1, regs, head);
536}
537
538int perf_sysenter_enable(struct ftrace_event_call *call)
539{
540 int ret = 0;
541 int num;
542
543 num = ((struct syscall_metadata *)call->data)->syscall_nr;
544
545 mutex_lock(&syscall_trace_lock);
546 if (!sys_perf_refcount_enter)
547 ret = register_trace_sys_enter(perf_syscall_enter, NULL);
548 if (ret) {
549 pr_info("event trace: Could not activate"
550 "syscall entry trace point");
551 } else {
552 set_bit(num, enabled_perf_enter_syscalls);
553 sys_perf_refcount_enter++;
554 }
555 mutex_unlock(&syscall_trace_lock);
556 return ret;
557}
558
559void perf_sysenter_disable(struct ftrace_event_call *call)
560{
561 int num;
562
563 num = ((struct syscall_metadata *)call->data)->syscall_nr;
564
565 mutex_lock(&syscall_trace_lock);
566 sys_perf_refcount_enter--;
567 clear_bit(num, enabled_perf_enter_syscalls);
568 if (!sys_perf_refcount_enter)
569 unregister_trace_sys_enter(perf_syscall_enter, NULL);
570 mutex_unlock(&syscall_trace_lock);
571}
572
573static void perf_syscall_exit(void *ignore, struct pt_regs *regs, long ret)
574{
575 struct syscall_metadata *sys_data;
576 struct syscall_trace_exit *rec;
577 struct hlist_head *head;
578 int syscall_nr;
579 int rctx;
580 int size;
581
582 syscall_nr = syscall_get_nr(current, regs);
583 if (!test_bit(syscall_nr, enabled_perf_exit_syscalls))
584 return;
585
586 sys_data = syscall_nr_to_meta(syscall_nr);
587 if (!sys_data)
588 return;
589
590 /* We can probably do that at build time */
591 size = ALIGN(sizeof(*rec) + sizeof(u32), sizeof(u64));
592 size -= sizeof(u32);
593
594 /*
595 * Impossible, but be paranoid with the future
596 * How to put this check outside runtime?
597 */
598 if (WARN_ONCE(size > PERF_MAX_TRACE_SIZE,
599 "exit event has grown above perf buffer size"))
600 return;
601
602 rec = (struct syscall_trace_exit *)perf_trace_buf_prepare(size,
603 sys_data->exit_event->event.type, regs, &rctx);
604 if (!rec)
605 return;
606
607 rec->nr = syscall_nr;
608 rec->ret = syscall_get_return_value(current, regs);
609
610 head = this_cpu_ptr(sys_data->exit_event->perf_events);
611 perf_trace_buf_submit(rec, size, rctx, 0, 1, regs, head);
612}
613
614int perf_sysexit_enable(struct ftrace_event_call *call)
615{
616 int ret = 0;
617 int num;
618
619 num = ((struct syscall_metadata *)call->data)->syscall_nr;
620
621 mutex_lock(&syscall_trace_lock);
622 if (!sys_perf_refcount_exit)
623 ret = register_trace_sys_exit(perf_syscall_exit, NULL);
624 if (ret) {
625 pr_info("event trace: Could not activate"
626 "syscall exit trace point");
627 } else {
628 set_bit(num, enabled_perf_exit_syscalls);
629 sys_perf_refcount_exit++;
630 }
631 mutex_unlock(&syscall_trace_lock);
632 return ret;
633}
634
635void perf_sysexit_disable(struct ftrace_event_call *call)
636{
637 int num;
638
639 num = ((struct syscall_metadata *)call->data)->syscall_nr;
640
641 mutex_lock(&syscall_trace_lock);
642 sys_perf_refcount_exit--;
643 clear_bit(num, enabled_perf_exit_syscalls);
644 if (!sys_perf_refcount_exit)
645 unregister_trace_sys_exit(perf_syscall_exit, NULL);
646 mutex_unlock(&syscall_trace_lock);
647}
648
649#endif /* CONFIG_PERF_EVENTS */
650
651static int syscall_enter_register(struct ftrace_event_call *event,
652 enum trace_reg type, void *data)
653{
654 switch (type) {
655 case TRACE_REG_REGISTER:
656 return reg_event_syscall_enter(event);
657 case TRACE_REG_UNREGISTER:
658 unreg_event_syscall_enter(event);
659 return 0;
660
661#ifdef CONFIG_PERF_EVENTS
662 case TRACE_REG_PERF_REGISTER:
663 return perf_sysenter_enable(event);
664 case TRACE_REG_PERF_UNREGISTER:
665 perf_sysenter_disable(event);
666 return 0;
667 case TRACE_REG_PERF_OPEN:
668 case TRACE_REG_PERF_CLOSE:
669 case TRACE_REG_PERF_ADD:
670 case TRACE_REG_PERF_DEL:
671 return 0;
672#endif
673 }
674 return 0;
675}
676
677static int syscall_exit_register(struct ftrace_event_call *event,
678 enum trace_reg type, void *data)
679{
680 switch (type) {
681 case TRACE_REG_REGISTER:
682 return reg_event_syscall_exit(event);
683 case TRACE_REG_UNREGISTER:
684 unreg_event_syscall_exit(event);
685 return 0;
686
687#ifdef CONFIG_PERF_EVENTS
688 case TRACE_REG_PERF_REGISTER:
689 return perf_sysexit_enable(event);
690 case TRACE_REG_PERF_UNREGISTER:
691 perf_sysexit_disable(event);
692 return 0;
693 case TRACE_REG_PERF_OPEN:
694 case TRACE_REG_PERF_CLOSE:
695 case TRACE_REG_PERF_ADD:
696 case TRACE_REG_PERF_DEL:
697 return 0;
698#endif
699 }
700 return 0;
701}
1#include <trace/syscall.h>
2#include <trace/events/syscalls.h>
3#include <linux/syscalls.h>
4#include <linux/slab.h>
5#include <linux/kernel.h>
6#include <linux/module.h> /* for MODULE_NAME_LEN via KSYM_SYMBOL_LEN */
7#include <linux/ftrace.h>
8#include <linux/perf_event.h>
9#include <asm/syscall.h>
10
11#include "trace_output.h"
12#include "trace.h"
13
14static DEFINE_MUTEX(syscall_trace_lock);
15
16static int syscall_enter_register(struct trace_event_call *event,
17 enum trace_reg type, void *data);
18static int syscall_exit_register(struct trace_event_call *event,
19 enum trace_reg type, void *data);
20
21static struct list_head *
22syscall_get_enter_fields(struct trace_event_call *call)
23{
24 struct syscall_metadata *entry = call->data;
25
26 return &entry->enter_fields;
27}
28
29extern struct syscall_metadata *__start_syscalls_metadata[];
30extern struct syscall_metadata *__stop_syscalls_metadata[];
31
32static struct syscall_metadata **syscalls_metadata;
33
34#ifndef ARCH_HAS_SYSCALL_MATCH_SYM_NAME
35static inline bool arch_syscall_match_sym_name(const char *sym, const char *name)
36{
37 /*
38 * Only compare after the "sys" prefix. Archs that use
39 * syscall wrappers may have syscalls symbols aliases prefixed
40 * with ".SyS" or ".sys" instead of "sys", leading to an unwanted
41 * mismatch.
42 */
43 return !strcmp(sym + 3, name + 3);
44}
45#endif
46
47#ifdef ARCH_TRACE_IGNORE_COMPAT_SYSCALLS
48/*
49 * Some architectures that allow for 32bit applications
50 * to run on a 64bit kernel, do not map the syscalls for
51 * the 32bit tasks the same as they do for 64bit tasks.
52 *
53 * *cough*x86*cough*
54 *
55 * In such a case, instead of reporting the wrong syscalls,
56 * simply ignore them.
57 *
58 * For an arch to ignore the compat syscalls it needs to
59 * define ARCH_TRACE_IGNORE_COMPAT_SYSCALLS as well as
60 * define the function arch_trace_is_compat_syscall() to let
61 * the tracing system know that it should ignore it.
62 */
63static int
64trace_get_syscall_nr(struct task_struct *task, struct pt_regs *regs)
65{
66 if (unlikely(arch_trace_is_compat_syscall(regs)))
67 return -1;
68
69 return syscall_get_nr(task, regs);
70}
71#else
72static inline int
73trace_get_syscall_nr(struct task_struct *task, struct pt_regs *regs)
74{
75 return syscall_get_nr(task, regs);
76}
77#endif /* ARCH_TRACE_IGNORE_COMPAT_SYSCALLS */
78
79static __init struct syscall_metadata *
80find_syscall_meta(unsigned long syscall)
81{
82 struct syscall_metadata **start;
83 struct syscall_metadata **stop;
84 char str[KSYM_SYMBOL_LEN];
85
86
87 start = __start_syscalls_metadata;
88 stop = __stop_syscalls_metadata;
89 kallsyms_lookup(syscall, NULL, NULL, NULL, str);
90
91 if (arch_syscall_match_sym_name(str, "sys_ni_syscall"))
92 return NULL;
93
94 for ( ; start < stop; start++) {
95 if ((*start)->name && arch_syscall_match_sym_name(str, (*start)->name))
96 return *start;
97 }
98 return NULL;
99}
100
101static struct syscall_metadata *syscall_nr_to_meta(int nr)
102{
103 if (!syscalls_metadata || nr >= NR_syscalls || nr < 0)
104 return NULL;
105
106 return syscalls_metadata[nr];
107}
108
109const char *get_syscall_name(int syscall)
110{
111 struct syscall_metadata *entry;
112
113 entry = syscall_nr_to_meta(syscall);
114 if (!entry)
115 return NULL;
116
117 return entry->name;
118}
119
120static enum print_line_t
121print_syscall_enter(struct trace_iterator *iter, int flags,
122 struct trace_event *event)
123{
124 struct trace_array *tr = iter->tr;
125 struct trace_seq *s = &iter->seq;
126 struct trace_entry *ent = iter->ent;
127 struct syscall_trace_enter *trace;
128 struct syscall_metadata *entry;
129 int i, syscall;
130
131 trace = (typeof(trace))ent;
132 syscall = trace->nr;
133 entry = syscall_nr_to_meta(syscall);
134
135 if (!entry)
136 goto end;
137
138 if (entry->enter_event->event.type != ent->type) {
139 WARN_ON_ONCE(1);
140 goto end;
141 }
142
143 trace_seq_printf(s, "%s(", entry->name);
144
145 for (i = 0; i < entry->nb_args; i++) {
146
147 if (trace_seq_has_overflowed(s))
148 goto end;
149
150 /* parameter types */
151 if (tr->trace_flags & TRACE_ITER_VERBOSE)
152 trace_seq_printf(s, "%s ", entry->types[i]);
153
154 /* parameter values */
155 trace_seq_printf(s, "%s: %lx%s", entry->args[i],
156 trace->args[i],
157 i == entry->nb_args - 1 ? "" : ", ");
158 }
159
160 trace_seq_putc(s, ')');
161end:
162 trace_seq_putc(s, '\n');
163
164 return trace_handle_return(s);
165}
166
167static enum print_line_t
168print_syscall_exit(struct trace_iterator *iter, int flags,
169 struct trace_event *event)
170{
171 struct trace_seq *s = &iter->seq;
172 struct trace_entry *ent = iter->ent;
173 struct syscall_trace_exit *trace;
174 int syscall;
175 struct syscall_metadata *entry;
176
177 trace = (typeof(trace))ent;
178 syscall = trace->nr;
179 entry = syscall_nr_to_meta(syscall);
180
181 if (!entry) {
182 trace_seq_putc(s, '\n');
183 goto out;
184 }
185
186 if (entry->exit_event->event.type != ent->type) {
187 WARN_ON_ONCE(1);
188 return TRACE_TYPE_UNHANDLED;
189 }
190
191 trace_seq_printf(s, "%s -> 0x%lx\n", entry->name,
192 trace->ret);
193
194 out:
195 return trace_handle_return(s);
196}
197
198extern char *__bad_type_size(void);
199
200#define SYSCALL_FIELD(type, field, name) \
201 sizeof(type) != sizeof(trace.field) ? \
202 __bad_type_size() : \
203 #type, #name, offsetof(typeof(trace), field), \
204 sizeof(trace.field), is_signed_type(type)
205
206static int __init
207__set_enter_print_fmt(struct syscall_metadata *entry, char *buf, int len)
208{
209 int i;
210 int pos = 0;
211
212 /* When len=0, we just calculate the needed length */
213#define LEN_OR_ZERO (len ? len - pos : 0)
214
215 pos += snprintf(buf + pos, LEN_OR_ZERO, "\"");
216 for (i = 0; i < entry->nb_args; i++) {
217 pos += snprintf(buf + pos, LEN_OR_ZERO, "%s: 0x%%0%zulx%s",
218 entry->args[i], sizeof(unsigned long),
219 i == entry->nb_args - 1 ? "" : ", ");
220 }
221 pos += snprintf(buf + pos, LEN_OR_ZERO, "\"");
222
223 for (i = 0; i < entry->nb_args; i++) {
224 pos += snprintf(buf + pos, LEN_OR_ZERO,
225 ", ((unsigned long)(REC->%s))", entry->args[i]);
226 }
227
228#undef LEN_OR_ZERO
229
230 /* return the length of print_fmt */
231 return pos;
232}
233
234static int __init set_syscall_print_fmt(struct trace_event_call *call)
235{
236 char *print_fmt;
237 int len;
238 struct syscall_metadata *entry = call->data;
239
240 if (entry->enter_event != call) {
241 call->print_fmt = "\"0x%lx\", REC->ret";
242 return 0;
243 }
244
245 /* First: called with 0 length to calculate the needed length */
246 len = __set_enter_print_fmt(entry, NULL, 0);
247
248 print_fmt = kmalloc(len + 1, GFP_KERNEL);
249 if (!print_fmt)
250 return -ENOMEM;
251
252 /* Second: actually write the @print_fmt */
253 __set_enter_print_fmt(entry, print_fmt, len + 1);
254 call->print_fmt = print_fmt;
255
256 return 0;
257}
258
259static void __init free_syscall_print_fmt(struct trace_event_call *call)
260{
261 struct syscall_metadata *entry = call->data;
262
263 if (entry->enter_event == call)
264 kfree(call->print_fmt);
265}
266
267static int __init syscall_enter_define_fields(struct trace_event_call *call)
268{
269 struct syscall_trace_enter trace;
270 struct syscall_metadata *meta = call->data;
271 int ret;
272 int i;
273 int offset = offsetof(typeof(trace), args);
274
275 ret = trace_define_field(call, SYSCALL_FIELD(int, nr, __syscall_nr),
276 FILTER_OTHER);
277 if (ret)
278 return ret;
279
280 for (i = 0; i < meta->nb_args; i++) {
281 ret = trace_define_field(call, meta->types[i],
282 meta->args[i], offset,
283 sizeof(unsigned long), 0,
284 FILTER_OTHER);
285 offset += sizeof(unsigned long);
286 }
287
288 return ret;
289}
290
291static int __init syscall_exit_define_fields(struct trace_event_call *call)
292{
293 struct syscall_trace_exit trace;
294 int ret;
295
296 ret = trace_define_field(call, SYSCALL_FIELD(int, nr, __syscall_nr),
297 FILTER_OTHER);
298 if (ret)
299 return ret;
300
301 ret = trace_define_field(call, SYSCALL_FIELD(long, ret, ret),
302 FILTER_OTHER);
303
304 return ret;
305}
306
307static void ftrace_syscall_enter(void *data, struct pt_regs *regs, long id)
308{
309 struct trace_array *tr = data;
310 struct trace_event_file *trace_file;
311 struct syscall_trace_enter *entry;
312 struct syscall_metadata *sys_data;
313 struct ring_buffer_event *event;
314 struct ring_buffer *buffer;
315 unsigned long irq_flags;
316 int pc;
317 int syscall_nr;
318 int size;
319
320 syscall_nr = trace_get_syscall_nr(current, regs);
321 if (syscall_nr < 0 || syscall_nr >= NR_syscalls)
322 return;
323
324 /* Here we're inside tp handler's rcu_read_lock_sched (__DO_TRACE) */
325 trace_file = rcu_dereference_sched(tr->enter_syscall_files[syscall_nr]);
326 if (!trace_file)
327 return;
328
329 if (trace_trigger_soft_disabled(trace_file))
330 return;
331
332 sys_data = syscall_nr_to_meta(syscall_nr);
333 if (!sys_data)
334 return;
335
336 size = sizeof(*entry) + sizeof(unsigned long) * sys_data->nb_args;
337
338 local_save_flags(irq_flags);
339 pc = preempt_count();
340
341 buffer = tr->trace_buffer.buffer;
342 event = trace_buffer_lock_reserve(buffer,
343 sys_data->enter_event->event.type, size, irq_flags, pc);
344 if (!event)
345 return;
346
347 entry = ring_buffer_event_data(event);
348 entry->nr = syscall_nr;
349 syscall_get_arguments(current, regs, 0, sys_data->nb_args, entry->args);
350
351 event_trigger_unlock_commit(trace_file, buffer, event, entry,
352 irq_flags, pc);
353}
354
355static void ftrace_syscall_exit(void *data, struct pt_regs *regs, long ret)
356{
357 struct trace_array *tr = data;
358 struct trace_event_file *trace_file;
359 struct syscall_trace_exit *entry;
360 struct syscall_metadata *sys_data;
361 struct ring_buffer_event *event;
362 struct ring_buffer *buffer;
363 unsigned long irq_flags;
364 int pc;
365 int syscall_nr;
366
367 syscall_nr = trace_get_syscall_nr(current, regs);
368 if (syscall_nr < 0 || syscall_nr >= NR_syscalls)
369 return;
370
371 /* Here we're inside tp handler's rcu_read_lock_sched (__DO_TRACE()) */
372 trace_file = rcu_dereference_sched(tr->exit_syscall_files[syscall_nr]);
373 if (!trace_file)
374 return;
375
376 if (trace_trigger_soft_disabled(trace_file))
377 return;
378
379 sys_data = syscall_nr_to_meta(syscall_nr);
380 if (!sys_data)
381 return;
382
383 local_save_flags(irq_flags);
384 pc = preempt_count();
385
386 buffer = tr->trace_buffer.buffer;
387 event = trace_buffer_lock_reserve(buffer,
388 sys_data->exit_event->event.type, sizeof(*entry),
389 irq_flags, pc);
390 if (!event)
391 return;
392
393 entry = ring_buffer_event_data(event);
394 entry->nr = syscall_nr;
395 entry->ret = syscall_get_return_value(current, regs);
396
397 event_trigger_unlock_commit(trace_file, buffer, event, entry,
398 irq_flags, pc);
399}
400
401static int reg_event_syscall_enter(struct trace_event_file *file,
402 struct trace_event_call *call)
403{
404 struct trace_array *tr = file->tr;
405 int ret = 0;
406 int num;
407
408 num = ((struct syscall_metadata *)call->data)->syscall_nr;
409 if (WARN_ON_ONCE(num < 0 || num >= NR_syscalls))
410 return -ENOSYS;
411 mutex_lock(&syscall_trace_lock);
412 if (!tr->sys_refcount_enter)
413 ret = register_trace_sys_enter(ftrace_syscall_enter, tr);
414 if (!ret) {
415 rcu_assign_pointer(tr->enter_syscall_files[num], file);
416 tr->sys_refcount_enter++;
417 }
418 mutex_unlock(&syscall_trace_lock);
419 return ret;
420}
421
422static void unreg_event_syscall_enter(struct trace_event_file *file,
423 struct trace_event_call *call)
424{
425 struct trace_array *tr = file->tr;
426 int num;
427
428 num = ((struct syscall_metadata *)call->data)->syscall_nr;
429 if (WARN_ON_ONCE(num < 0 || num >= NR_syscalls))
430 return;
431 mutex_lock(&syscall_trace_lock);
432 tr->sys_refcount_enter--;
433 RCU_INIT_POINTER(tr->enter_syscall_files[num], NULL);
434 if (!tr->sys_refcount_enter)
435 unregister_trace_sys_enter(ftrace_syscall_enter, tr);
436 mutex_unlock(&syscall_trace_lock);
437}
438
439static int reg_event_syscall_exit(struct trace_event_file *file,
440 struct trace_event_call *call)
441{
442 struct trace_array *tr = file->tr;
443 int ret = 0;
444 int num;
445
446 num = ((struct syscall_metadata *)call->data)->syscall_nr;
447 if (WARN_ON_ONCE(num < 0 || num >= NR_syscalls))
448 return -ENOSYS;
449 mutex_lock(&syscall_trace_lock);
450 if (!tr->sys_refcount_exit)
451 ret = register_trace_sys_exit(ftrace_syscall_exit, tr);
452 if (!ret) {
453 rcu_assign_pointer(tr->exit_syscall_files[num], file);
454 tr->sys_refcount_exit++;
455 }
456 mutex_unlock(&syscall_trace_lock);
457 return ret;
458}
459
460static void unreg_event_syscall_exit(struct trace_event_file *file,
461 struct trace_event_call *call)
462{
463 struct trace_array *tr = file->tr;
464 int num;
465
466 num = ((struct syscall_metadata *)call->data)->syscall_nr;
467 if (WARN_ON_ONCE(num < 0 || num >= NR_syscalls))
468 return;
469 mutex_lock(&syscall_trace_lock);
470 tr->sys_refcount_exit--;
471 RCU_INIT_POINTER(tr->exit_syscall_files[num], NULL);
472 if (!tr->sys_refcount_exit)
473 unregister_trace_sys_exit(ftrace_syscall_exit, tr);
474 mutex_unlock(&syscall_trace_lock);
475}
476
477static int __init init_syscall_trace(struct trace_event_call *call)
478{
479 int id;
480 int num;
481
482 num = ((struct syscall_metadata *)call->data)->syscall_nr;
483 if (num < 0 || num >= NR_syscalls) {
484 pr_debug("syscall %s metadata not mapped, disabling ftrace event\n",
485 ((struct syscall_metadata *)call->data)->name);
486 return -ENOSYS;
487 }
488
489 if (set_syscall_print_fmt(call) < 0)
490 return -ENOMEM;
491
492 id = trace_event_raw_init(call);
493
494 if (id < 0) {
495 free_syscall_print_fmt(call);
496 return id;
497 }
498
499 return id;
500}
501
502struct trace_event_functions enter_syscall_print_funcs = {
503 .trace = print_syscall_enter,
504};
505
506struct trace_event_functions exit_syscall_print_funcs = {
507 .trace = print_syscall_exit,
508};
509
510struct trace_event_class __refdata event_class_syscall_enter = {
511 .system = "syscalls",
512 .reg = syscall_enter_register,
513 .define_fields = syscall_enter_define_fields,
514 .get_fields = syscall_get_enter_fields,
515 .raw_init = init_syscall_trace,
516};
517
518struct trace_event_class __refdata event_class_syscall_exit = {
519 .system = "syscalls",
520 .reg = syscall_exit_register,
521 .define_fields = syscall_exit_define_fields,
522 .fields = LIST_HEAD_INIT(event_class_syscall_exit.fields),
523 .raw_init = init_syscall_trace,
524};
525
526unsigned long __init __weak arch_syscall_addr(int nr)
527{
528 return (unsigned long)sys_call_table[nr];
529}
530
531void __init init_ftrace_syscalls(void)
532{
533 struct syscall_metadata *meta;
534 unsigned long addr;
535 int i;
536
537 syscalls_metadata = kcalloc(NR_syscalls, sizeof(*syscalls_metadata),
538 GFP_KERNEL);
539 if (!syscalls_metadata) {
540 WARN_ON(1);
541 return;
542 }
543
544 for (i = 0; i < NR_syscalls; i++) {
545 addr = arch_syscall_addr(i);
546 meta = find_syscall_meta(addr);
547 if (!meta)
548 continue;
549
550 meta->syscall_nr = i;
551 syscalls_metadata[i] = meta;
552 }
553}
554
555#ifdef CONFIG_PERF_EVENTS
556
557static DECLARE_BITMAP(enabled_perf_enter_syscalls, NR_syscalls);
558static DECLARE_BITMAP(enabled_perf_exit_syscalls, NR_syscalls);
559static int sys_perf_refcount_enter;
560static int sys_perf_refcount_exit;
561
562static void perf_syscall_enter(void *ignore, struct pt_regs *regs, long id)
563{
564 struct syscall_metadata *sys_data;
565 struct syscall_trace_enter *rec;
566 struct hlist_head *head;
567 int syscall_nr;
568 int rctx;
569 int size;
570
571 syscall_nr = trace_get_syscall_nr(current, regs);
572 if (syscall_nr < 0 || syscall_nr >= NR_syscalls)
573 return;
574 if (!test_bit(syscall_nr, enabled_perf_enter_syscalls))
575 return;
576
577 sys_data = syscall_nr_to_meta(syscall_nr);
578 if (!sys_data)
579 return;
580
581 head = this_cpu_ptr(sys_data->enter_event->perf_events);
582 if (hlist_empty(head))
583 return;
584
585 /* get the size after alignment with the u32 buffer size field */
586 size = sizeof(unsigned long) * sys_data->nb_args + sizeof(*rec);
587 size = ALIGN(size + sizeof(u32), sizeof(u64));
588 size -= sizeof(u32);
589
590 rec = perf_trace_buf_alloc(size, NULL, &rctx);
591 if (!rec)
592 return;
593
594 rec->nr = syscall_nr;
595 syscall_get_arguments(current, regs, 0, sys_data->nb_args,
596 (unsigned long *)&rec->args);
597 perf_trace_buf_submit(rec, size, rctx,
598 sys_data->enter_event->event.type, 1, regs,
599 head, NULL);
600}
601
602static int perf_sysenter_enable(struct trace_event_call *call)
603{
604 int ret = 0;
605 int num;
606
607 num = ((struct syscall_metadata *)call->data)->syscall_nr;
608
609 mutex_lock(&syscall_trace_lock);
610 if (!sys_perf_refcount_enter)
611 ret = register_trace_sys_enter(perf_syscall_enter, NULL);
612 if (ret) {
613 pr_info("event trace: Could not activate syscall entry trace point");
614 } else {
615 set_bit(num, enabled_perf_enter_syscalls);
616 sys_perf_refcount_enter++;
617 }
618 mutex_unlock(&syscall_trace_lock);
619 return ret;
620}
621
622static void perf_sysenter_disable(struct trace_event_call *call)
623{
624 int num;
625
626 num = ((struct syscall_metadata *)call->data)->syscall_nr;
627
628 mutex_lock(&syscall_trace_lock);
629 sys_perf_refcount_enter--;
630 clear_bit(num, enabled_perf_enter_syscalls);
631 if (!sys_perf_refcount_enter)
632 unregister_trace_sys_enter(perf_syscall_enter, NULL);
633 mutex_unlock(&syscall_trace_lock);
634}
635
636static void perf_syscall_exit(void *ignore, struct pt_regs *regs, long ret)
637{
638 struct syscall_metadata *sys_data;
639 struct syscall_trace_exit *rec;
640 struct hlist_head *head;
641 int syscall_nr;
642 int rctx;
643 int size;
644
645 syscall_nr = trace_get_syscall_nr(current, regs);
646 if (syscall_nr < 0 || syscall_nr >= NR_syscalls)
647 return;
648 if (!test_bit(syscall_nr, enabled_perf_exit_syscalls))
649 return;
650
651 sys_data = syscall_nr_to_meta(syscall_nr);
652 if (!sys_data)
653 return;
654
655 head = this_cpu_ptr(sys_data->exit_event->perf_events);
656 if (hlist_empty(head))
657 return;
658
659 /* We can probably do that at build time */
660 size = ALIGN(sizeof(*rec) + sizeof(u32), sizeof(u64));
661 size -= sizeof(u32);
662
663 rec = perf_trace_buf_alloc(size, NULL, &rctx);
664 if (!rec)
665 return;
666
667 rec->nr = syscall_nr;
668 rec->ret = syscall_get_return_value(current, regs);
669 perf_trace_buf_submit(rec, size, rctx, sys_data->exit_event->event.type,
670 1, regs, head, NULL);
671}
672
673static int perf_sysexit_enable(struct trace_event_call *call)
674{
675 int ret = 0;
676 int num;
677
678 num = ((struct syscall_metadata *)call->data)->syscall_nr;
679
680 mutex_lock(&syscall_trace_lock);
681 if (!sys_perf_refcount_exit)
682 ret = register_trace_sys_exit(perf_syscall_exit, NULL);
683 if (ret) {
684 pr_info("event trace: Could not activate syscall exit trace point");
685 } else {
686 set_bit(num, enabled_perf_exit_syscalls);
687 sys_perf_refcount_exit++;
688 }
689 mutex_unlock(&syscall_trace_lock);
690 return ret;
691}
692
693static void perf_sysexit_disable(struct trace_event_call *call)
694{
695 int num;
696
697 num = ((struct syscall_metadata *)call->data)->syscall_nr;
698
699 mutex_lock(&syscall_trace_lock);
700 sys_perf_refcount_exit--;
701 clear_bit(num, enabled_perf_exit_syscalls);
702 if (!sys_perf_refcount_exit)
703 unregister_trace_sys_exit(perf_syscall_exit, NULL);
704 mutex_unlock(&syscall_trace_lock);
705}
706
707#endif /* CONFIG_PERF_EVENTS */
708
709static int syscall_enter_register(struct trace_event_call *event,
710 enum trace_reg type, void *data)
711{
712 struct trace_event_file *file = data;
713
714 switch (type) {
715 case TRACE_REG_REGISTER:
716 return reg_event_syscall_enter(file, event);
717 case TRACE_REG_UNREGISTER:
718 unreg_event_syscall_enter(file, event);
719 return 0;
720
721#ifdef CONFIG_PERF_EVENTS
722 case TRACE_REG_PERF_REGISTER:
723 return perf_sysenter_enable(event);
724 case TRACE_REG_PERF_UNREGISTER:
725 perf_sysenter_disable(event);
726 return 0;
727 case TRACE_REG_PERF_OPEN:
728 case TRACE_REG_PERF_CLOSE:
729 case TRACE_REG_PERF_ADD:
730 case TRACE_REG_PERF_DEL:
731 return 0;
732#endif
733 }
734 return 0;
735}
736
737static int syscall_exit_register(struct trace_event_call *event,
738 enum trace_reg type, void *data)
739{
740 struct trace_event_file *file = data;
741
742 switch (type) {
743 case TRACE_REG_REGISTER:
744 return reg_event_syscall_exit(file, event);
745 case TRACE_REG_UNREGISTER:
746 unreg_event_syscall_exit(file, event);
747 return 0;
748
749#ifdef CONFIG_PERF_EVENTS
750 case TRACE_REG_PERF_REGISTER:
751 return perf_sysexit_enable(event);
752 case TRACE_REG_PERF_UNREGISTER:
753 perf_sysexit_disable(event);
754 return 0;
755 case TRACE_REG_PERF_OPEN:
756 case TRACE_REG_PERF_CLOSE:
757 case TRACE_REG_PERF_ADD:
758 case TRACE_REG_PERF_DEL:
759 return 0;
760#endif
761 }
762 return 0;
763}